\documentclass[12pt,]{article}
\usepackage[]{mathpazo}
\usepackage{amssymb,amsmath}
\usepackage{ifxetex,ifluatex}
\usepackage{fixltx2e} % provides \textsubscript
\ifnum 0\ifxetex 1\fi\ifluatex 1\fi=0 % if pdftex
  \usepackage[T1]{fontenc}
  \usepackage[utf8]{inputenc}
\else % if luatex or xelatex
  \ifxetex
    \usepackage{mathspec}
  \else
    \usepackage{fontspec}
  \fi
  \defaultfontfeatures{Ligatures=TeX,Scale=MatchLowercase}
\fi
% use upquote if available, for straight quotes in verbatim environments
\IfFileExists{upquote.sty}{\usepackage{upquote}}{}
% use microtype if available
\IfFileExists{microtype.sty}{%
\usepackage{microtype}
\UseMicrotypeSet[protrusion]{basicmath} % disable protrusion for tt fonts
}{}
\usepackage[margin=1in]{geometry}
\usepackage{hyperref}
\hypersetup{unicode=true,
            pdftitle={Online Supporting Information},
            pdfauthor={Gabriel Lenz; Alexander Sahn},
            pdfborder={0 0 0},
            breaklinks=true}
\urlstyle{same}  % don't use monospace font for urls
\usepackage{natbib}
\bibliographystyle{apsr}
\usepackage{longtable,booktabs}
\usepackage{graphicx,grffile}
\makeatletter
\def\maxwidth{\ifdim\Gin@nat@width>\linewidth\linewidth\else\Gin@nat@width\fi}
\def\maxheight{\ifdim\Gin@nat@height>\textheight\textheight\else\Gin@nat@height\fi}
\makeatother
% Scale images if necessary, so that they will not overflow the page
% margins by default, and it is still possible to overwrite the defaults
% using explicit options in \includegraphics[width, height, ...]{}
\setkeys{Gin}{width=\maxwidth,height=\maxheight,keepaspectratio}
\IfFileExists{parskip.sty}{%
\usepackage{parskip}
}{% else
\setlength{\parindent}{0pt}
\setlength{\parskip}{6pt plus 2pt minus 1pt}
}
\setlength{\emergencystretch}{3em}  % prevent overfull lines
\providecommand{\tightlist}{%
  \setlength{\itemsep}{0pt}\setlength{\parskip}{0pt}}
\setcounter{secnumdepth}{0}
% Redefines (sub)paragraphs to behave more like sections
\ifx\paragraph\undefined\else
\let\oldparagraph\paragraph
\renewcommand{\paragraph}[1]{\oldparagraph{#1}\mbox{}}
\fi
\ifx\subparagraph\undefined\else
\let\oldsubparagraph\subparagraph
\renewcommand{\subparagraph}[1]{\oldsubparagraph{#1}\mbox{}}
\fi

%%% Use protect on footnotes to avoid problems with footnotes in titles
\let\rmarkdownfootnote\footnote%
\def\footnote{\protect\rmarkdownfootnote}

%%% Change title format to be more compact
\usepackage{titling}

% Create subtitle command for use in maketitle
\providecommand{\subtitle}[1]{
  \posttitle{
    \begin{center}\large#1\end{center}
    }
}

\setlength{\droptitle}{-2em}

  \title{Online Supporting Information}
    \pretitle{\vspace{\droptitle}\centering\huge}
  \posttitle{\par}
  \subtitle{Achieving Statistical Significance with Control Variables and without
Transparency}
  \author{Gabriel Lenz \\ Alexander Sahn}
    \preauthor{\centering\large\emph}
  \postauthor{\par}
      \predate{\centering\large\emph}
  \postdate{\par}
    \date{June 01, 2020}


\begin{document}
\maketitle

{
\setcounter{tocdepth}{2}
\tableofcontents
}
\newpage

\hypertarget{quotes-on-suppression-effects}{%
\section{Quotes on Suppression
Effects}\label{quotes-on-suppression-effects}}

``Ideally, including suppressor variables in a model should be theory
based and every regression model should include using a test for
suppressor effects. This approach allows researchers to become ware of
the suppressor effect of a particular variable and to be better able to
explain when regression results change drastically from one model to
another.'' \citep{pandey_suppressor_2010}

``Suppression and multicollinearity present challenging conditions for
those who wish to apply multiple regression analysis.''
\citep{beckstead_isolating_2012}

``The authors discuss definitions of the suppressor phenomenon, show how
the unwary researcher can be warned against it, and present guidelines
for the interpretation of the results.'' \citep{maassen_suppressor_2001}

``In his influential book, Personality and Prediction, Jerry Wiggins
(1973) concluded:''the case for suppressor variables remains to be
demonstrated" (p.~38). The impact of the book on a generation of
personality researchers led them, quite rightly, to be skeptical about
the value of suppressor variables. Wiggins's conclusion was based partly
on Ghiselli's (1972) failure to replicate a suppressor effect, leading
him to liken suppressor variables to the ephemeral `will-o-the wisp'
(p.~270). Wiggins was particularly persuaded by the disappointing
results reported by Goldberg, Rorer, and Green (1970). Since those
earlier warnings, summary evaluations of the utility of suppressor
variables in behavioral science have remained guarded at best (e.g.,
Cohen \& Cohen, 1992; Pedhazur, 1982). Even very recently, Maassen and
Baker (2001) warned against devoting energy to formulating theoretical
explanations for solitary suppression results.''
\citep{paulhus_two_2004}

"The most difficult interpretive problem, however, occurs in the face of
suppression when relations between an independent variable and a
dependent variable increase or change direction following partialling
(e.g., Cohen \& Cohen, 1983; Darlington, 1968; Horst, 1941). This
situation poses the greatest interpretive hazard because a relation that
did not exist or did not exist as strongly before partialling is now
uncovered'' \citep{lynam_perils_2006}.

"Suppression effects in multiple linear regression are one of the most
elusive phenomena in the educational and psychological measurement
literature.'' \citep{kim_causal_2019}

We note that some of these authors are using suppression effects to
describe a very particular case where the suppressor variable is
uncorrelated with the dependent variable.

\hypertarget{quotes-on-controls-introducing-bias}{%
\subsection{Quotes on Controls Introducing
Bias}\label{quotes-on-controls-introducing-bias}}

``Omitting both the fixed effects and the unobserved confounder was
preferable to adjusting for the fixed effects precisely because the two
biases counterbalanced one another in the unadjusted estimate. In
practice, a researcher is unlikely to know whether adjusting for
covariates will unmask unobserved confounder bias. Similar observations
have led to somewhat pessimistic assessment of observational analysis,
for example, in Clarke (2005) and Frisell et al. (2012) (but see also
Clarke {[}2009{]}).'' \citep{middleton_bias_2016}

``Using a simple linear (regression) setting with two confounders one
observed (X), the other unobserved (U) we demonstrate that conditioning
on the observed confounder X does not necessarily imply that the
confounding bias decreases, even if X is highly correlated with U. That
is, adjusting for X may increase instead of reduce the omitted variable
bias (OVB). Two phenomena can cause an increasing OVB: (i) bias
amplification and (ii) cancellation of offsetting biases.''
\citep{steiner_mechanics_2016}

``A key underlying assumption is that the danger posed by omitted
variable bias can be ameliorated by the inclusion of relevant control
variables. Unfortunately, as this article demonstrates, there is nothing
in the mathematics of regression analysis that supports this conclusion.
The inclusion of additional control variables may increase or decrease
the bias, and we cannot know for sure which is the case in any
particular situation.'' \citep{clarke_phantom_2005}

``Scholars often assume that the danger posed by omitted variable bias
can be ameliorated by the inclusion of large numbers of relevant control
variables. However, there is nothing in the mathematics of regression
analysis that supports this conclusion. This paper goes beyond textbook
treatments of omitted variable bias and shows, both for OLS and for
generalized linear models, that the inclusion of additional control
variables may increase or decrease the bias, and we cannot know for sure
which is the case in any particular situation. The last section of the
paper shows how formal sensitivity analysis can be used to determine
whether omitted variables are a problem. A substantive example
demonstrates the method.'' \citep{clarke_return_2009}

\newpage

\hypertarget{method}{%
\section{Method}\label{method}}

According to our pre-analysis plan (see anonymous link:
\url{https://goo.gl/tCeNFM}), we selected articles and models from
within these articles according to the following rules. We mark aspects
of our plan that evolved as we collected the data in italics.

\hypertarget{article-selection-rules}{%
\subsection{Article Selection Rules}\label{article-selection-rules}}

We included \emph{AJPS} articles from volumes 57-59, because the strict
replication data and code posting was fully in effect for publish
articles by volume 57.

We selected articles according to the following rules:

\begin{itemize}
\tightlist
\item
  The article must rely on a statistical model (excludes experiments
  without controls, formal models, political theory, simulations, etc.).
\item
  The article's analysis uses \emph{a standard statistical model} and
  replication code and data is available.
\item
  The article's analysis uses three or more controls.

  \begin{itemize}
  \tightlist
  \item
    \emph{To increase the number of studies, we lowered the threshold to
    just one control.}
  \end{itemize}
\item
  The article's publication rest primary on a single, non-null result
  (excludes findings driven by multiple results or null results).
\end{itemize}

Table S1 presents a count of the articles excluded by the reason for
exclusion.

\hypertarget{model-selection-rules}{%
\subsection{Model Selection Rules}\label{model-selection-rules}}

Given that papers often have multiple hypotheses and multiple tests of
each hypothesis, we selected one statistical model from each paper
according to the following criteria:

\begin{itemize}
\tightlist
\item
  We selected a model that is the main finding of the paper-referenced
  in the abstract or in body.
\item
  If there are multiple specifications of the same model with different
  sets of controls, we chose the model with the full set of controls.
\item
  If there are multiple cases or measures of the key independent
  variable and the text does not point to one specifically, \emph{we
  used the summary measure, if the paper presents one, and if not,} we
  used the model with the smallest absolute value coefficient on the key
  variable.
\item
  We included \emph{any variables necessary for the research design or
  identification strategy}.
\end{itemize}

\newpage

\hypertarget{table-s1-reasons-for-exclusion}{%
\section{Table S1: Reasons for
Exclusion}\label{table-s1-reasons-for-exclusion}}

\begin{longtable}[]{@{}lr@{}}
\toprule
\endhead
Data or code unavailable & 6\tabularnewline
Formal model & 19\tabularnewline
Measurement focus & 3\tabularnewline
No control variables & 35\tabularnewline
Nonstandard statistical model & 13\tabularnewline
Null finding & 2\tabularnewline
Plethora of claims & 14\tabularnewline
Simulation & 2\tabularnewline
Theory & 5\tabularnewline
Total articles excluded & 99\tabularnewline
\bottomrule
\end{longtable}

``Nonstandard statistical model'' are those without a function in Stata
or R.

``Plethora of claims'' describes articles that did not depend on one key
statistically significant result, but on several results.

We exclude all AJPS ``workshop'' articles.

\newpage

\hypertarget{figure-s1-p-value-changes-in-all-studies}{%
\section{Figure S1: P-value Changes in All
Studies}\label{figure-s1-p-value-changes-in-all-studies}}

In the plot below, we show the average change in p-values from the
bivariate to the full specification for \emph{all} articles, not just
those that do not disclose the bivariate. The figure first shows p-value
changes all studies, observational and experimental, with 95\%
confidence intervals. It next shows the key finding in the paper, that
observational studies showing a bivariate specification have only a
trivial amount of p-value decreases, but that observational studies that
do not show a bivariate specification have a larger amount. Next, it
shows that the p-value decreases occur in studies with fixed effects and
without fixed effects. It then shows the p-value changes for all three
\emph{AJPS} volumes we analyze and all three subfields. Finally, it
shows estimates with below the median number of controls and above the
median number of controls. The mean number of controls in these studies
is nine and the median is eight. In the paper, we show similar
statistics only for studies that did not to show a bivariate
specification.

~~

\includegraphics{covariates_appendix_files/figure-latex/unnamed-chunk-2-1.pdf}

\newpage

\hypertarget{figure-s2-p-value-decreases-from-coefficient-change-only}{%
\section{Figure S2: P-value Decreases from Coefficient Change
Only}\label{figure-s2-p-value-decreases-from-coefficient-change-only}}

In the plot below, we show the average change in p-values from the
bivariate to the full specification (with 95\% confidence intervals) for
observational and experimental studies due only to coefficient change.
The figure shows that the majority of the p-value change shown in the
figures above are due to increases in the magnitude of the coefficient
rather than a reduction in the standard errors. This finding is robust
across different subsets of studies.

~~

\includegraphics{covariates_appendix_files/figure-latex/unnamed-chunk-3-1.pdf}

\newpage

\hypertarget{figure-s3-control-effects-in-bivariate-versus-full-specifications}{%
\section{Figure S3: Control Effects in Bivariate Versus Full
Specifications}\label{figure-s3-control-effects-in-bivariate-versus-full-specifications}}

As we discuss in the paper, one potential objection to our argument is
that researchers can't justify/explain suppression effects because of
the complexity of the multivariate space. This figure shows otherwise.
It reveals that controls generally have similar effects on the key
coefficient estimate in the simple bivariate case and in the complex
multivariate case. Each point represents a control in a regression
model, with the x value representing the effect of removing that control
from the full specification and the y value representing the effect of
adding it to the bivariate specification. All effects are standardized
to a mean of zero and standard deviation of one. The dotted line shows
the 45 degree line. The figure shows a strong relationship between these
two, implying that explanations about the effect of controls on key
effect estimates often translate from the bivariate to the full
specification.

\includegraphics{covariates_appendix_files/figure-latex/unnamed-chunk-4-1.pdf}

\newpage

\hypertarget{figure-s4-changes-in-coefficients-across-all-studies}{%
\section{Figure S4: Changes in Coefficients across All
Studies}\label{figure-s4-changes-in-coefficients-across-all-studies}}

This figure is analogous to S1 but for coefficient changes instead of
p-value changes. We calculate coefficient percent changes with log plus
one. ~~

\includegraphics{covariates_appendix_files/figure-latex/unnamed-chunk-5-1.pdf}

\newpage

\hypertarget{figure-s5-changes-in-coefficients-that-do-not-show-bivariate}{%
\section{Figure S5: Changes in Coefficients that Do Not Show
Bivariate}\label{figure-s5-changes-in-coefficients-that-do-not-show-bivariate}}

This figure is analogous to S2 but for coefficient changes instead of
p-value changes. We calculate coefficient percentchanges with log plus
one. ~~

\includegraphics{covariates_appendix_files/figure-latex/unnamed-chunk-6-1.pdf}

\newpage

\hypertarget{figure-s6-t-value-changes-in-observational-studies}{%
\section{Figure S6: T-Value Changes in Observational
Studies}\label{figure-s6-t-value-changes-in-observational-studies}}

\includegraphics{covariates_appendix_files/figure-latex/unnamed-chunk-7-1.pdf}

\newpage

\hypertarget{figure-s7a-confidence-intervals-of-bivariate-and-full-specifications}{%
\section{Figure S7a: Confidence Intervals of Bivariate and Full
Specifications}\label{figure-s7a-confidence-intervals-of-bivariate-and-full-specifications}}

\includegraphics{covariates_appendix_files/figure-latex/unnamed-chunk-8-1.pdf}

\newpage

\hypertarget{figure-s7b-confidence-intervals-of-bivariate-and-full-specifications-normed-absolute-value-coefficients}{%
\section{Figure S7b: Confidence Intervals of Bivariate and Full
Specifications (Normed Absolute Value
Coefficients)}\label{figure-s7b-confidence-intervals-of-bivariate-and-full-specifications-normed-absolute-value-coefficients}}

Negative vales are from cases that switch signs. All variables recoded
to 0-1.

\includegraphics{covariates_appendix_files/figure-latex/unnamed-chunk-9-1.pdf}

\newpage

\hypertarget{figure-s7c-confidence-intervals-of-bivariate-and-full-specifications-absolute-value-beta-coefficients}{%
\section{Figure S7c: Confidence Intervals of Bivariate and Full
Specifications (Absolute Value Beta
Coefficients)}\label{figure-s7c-confidence-intervals-of-bivariate-and-full-specifications-absolute-value-beta-coefficients}}

All variables standardized so that their variances are 1. Negative vales
are from cases that switch signs.

\includegraphics{covariates_appendix_files/figure-latex/unnamed-chunk-10-1.pdf}

\newpage

\hypertarget{figure-s7d-confidence-intervals-of-bivariate-and-full-specifications-absolute-value-beta-coefficientszoomed}{%
\section{Figure S7d: Confidence Intervals of Bivariate and Full
Specifications (Absolute Value Beta
Coefficients--Zoomed)}\label{figure-s7d-confidence-intervals-of-bivariate-and-full-specifications-absolute-value-beta-coefficientszoomed}}

All variables standardized so that their variances are 1. Negative vales
are from cases that switch signs.

\includegraphics{covariates_appendix_files/figure-latex/unnamed-chunk-11-1.pdf}

\newpage

\hypertarget{figure-s8-multicollinearity-in-observational-studies}{%
\section{Figure S8: Multicollinearity in Observational
Studies}\label{figure-s8-multicollinearity-in-observational-studies}}

The figure below shows the correlation between the key variable of
interest and each control variable for every observational study. This
figure follows the same design as Figures 2 and 3, which present our
main findings, including the exact same ordering of studies. As we
report in the paper, we find suppression effects helping authors achieve
statistical significance on the far right side of this figure. If
collinearity was contributing, we would expect to see higher
correlations on the far right of this figure. As the figure shows,
however, we see no such pattern.

\includegraphics{covariates_appendix_files/figure-latex/unnamed-chunk-12-1.pdf}

\newpage

\hypertarget{specification-uncertainty}{%
\section{Specification Uncertainty}\label{specification-uncertainty}}

We conducted analyses to assess specification uncertainty from control
choice, but concluded that these analyses were not especially
informative. Here we present two of these analyses.

\hypertarget{statistically-significant-in-all-specifications}{%
\subsection{Statistically Significant in All
Specifications}\label{statistically-significant-in-all-specifications}}

In the figure below, we present for each article the percent of all
possible control combinations in which the key effect estimate was
statistically significant (at the 0.05 level). Each dot represents one
study and the studies are sorted by percent significant. The figure
reveals that about a third of the studies were significant across all
specifications. The other two thirds vary considerably with some being
rarely statistically significant or never statistically significant. In
studies with a large number of controls, we randomly selected control
combinations.

~

\includegraphics{covariates_appendix_files/figure-latex/unnamed-chunk-13-1.pdf}

\newpage

\hypertarget{standard-deviation-of-estimates}{%
\subsection{Standard Deviation of
Estimates}\label{standard-deviation-of-estimates}}

In the figure below, we present the standard deviation of the key
variable's coefficient estimate across all possible control
combinations. To scale them, we present a ratio: this standard deviation
over the standard error of the key estimate (from the full
specification). This ratio captures the degree to which specification
uncertainty (from control choice) is larger or smaller than uncertainty
reported by the articles' standard errors on the key estimates. Ratios
of one, which are noted by the dotted line in the figure, show articles
where the standard deviation across all possible estimates is about the
same size as the standard error. If one believes that this standard
deviation captures specification uncertainty, then ratios of one also
imply that findings have about twice the uncertainty they report in
their standard errors (see the previous page for why you should probably
not draw this conclusion). The figure shows several articles that have
truly massive amounts of uncertainty relative to the standard errors
they report (as much as ninefold), but more than half have substantially
less uncertainty from control choice than reflected in their standard
errors.

~

\includegraphics{covariates_appendix_files/figure-latex/unnamed-chunk-14-1.pdf}

\newpage

\hypertarget{analyses-specified-in-pre-analysis-plan}{%
\section{Analyses Specified in Pre-Analysis
Plan}\label{analyses-specified-in-pre-analysis-plan}}

\hypertarget{inflating-with-controls}{%
\subsection{Inflating with Controls}\label{inflating-with-controls}}

\hypertarget{plots-of-key-coefficient-under-full-and-bivariate-specifications}{%
\subsubsection{Plots of Key Coefficient under Full and Bivariate
Specifications}\label{plots-of-key-coefficient-under-full-and-bivariate-specifications}}

The first of set of statistics we will estimate is straightforward. How
much does the inclusion of the controls used in researchers' main model
change the main finding, compared to a model without controls? We will
examine both the effect size estimate and the p-value. Often,
researchers never report the estimate without controls (and we will code
how often they do). This comparison will give us an overall sense of how
often controls inflate estimates.

Since we did not know what the variation of the coefficients would look
like, we prespecified three plots comparing the bivariate (often
bivariate) and multivariate coefficients for each paper

\begin{itemize}
\tightlist
\item
  The raw coefficients
\item
  standardized beta coefficients, and
\item
  the coefficients after rescaling variables to vary between zero and
  one.
\end{itemize}

The next three plots show those results. Unfortunately, the huge range
and scales makes them almost impossible to interpret. \newpage

\hypertarget{key-estimate-under-full-and-bivariate-specifications}{%
\subsubsection{Key Estimate Under Full and Bivariate
Specifications}\label{key-estimate-under-full-and-bivariate-specifications}}

\includegraphics{covariates_appendix_files/figure-latex/unnamed-chunk-15-1.pdf}
\newpage

\hypertarget{key-estimate-under-full-and-bivariate-specifications-beta-coefficients}{%
\subsubsection{Key Estimate Under Full and Bivariate Specifications
(Beta
Coefficients)}\label{key-estimate-under-full-and-bivariate-specifications-beta-coefficients}}

\includegraphics{covariates_appendix_files/figure-latex/unnamed-chunk-16-1.pdf}
\newpage

\hypertarget{key-estimate-under-full-and-bivariate-specifications-normed-coefficients}{%
\subsubsection{Key Estimate Under Full and Bivariate Specifications
(Normed
Coefficients)}\label{key-estimate-under-full-and-bivariate-specifications-normed-coefficients}}

\includegraphics{covariates_appendix_files/figure-latex/unnamed-chunk-17-1.pdf}
\newpage

\hypertarget{descriptive-statistics-on-inflation}{%
\subsubsection{Descriptive Statistics on
Inflation}\label{descriptive-statistics-on-inflation}}

\begin{itemize}
\item
  Percent of studies where the coefficient of the key estimate under
  full specification inflates over the bivariate specification: 53\% (34
  of 64). Note that this includes all studies. When limited to studies
  where a bivariate specification is not shown, the percentage is 64\%.
  When further limited to observational studies that do not show a
  bivariate specification, the percentage is 68\%. When further limited
  to studies whose bivariate specification had a p-value greater than
  0.05, the percentage is 100\%.
\item
  Percent of studies where p-value of key estimate does not achieve
  statistical significance under bivariate specification, but does in
  full specification: 25\% (16 of 64). 69\% of bivariate specification
  key estimates achieve statistical significance, 91\% of full
  specification key estimates achieve statistical significance at
  p\textless{}0.05).
\item
  Percent of models where the sign on the key estimate is different in
  the full and bivariate specifications 6\% (2\% where both estimates
  are statistically significant).
\end{itemize}

\hypertarget{sensitivity-to-control-choice}{%
\subsection{Sensitivity to Control
Choice}\label{sensitivity-to-control-choice}}

\begin{itemize}
\tightlist
\item
  98\% of alternative specifications have a key effect coefficient of
  the same sign as the full specification.
\item
  74\% of alternative specifications have a key effect coefficient of
  the same sign as the full specification that is statistically
  significant.
\item
  56\% of alternative specifications have a smaller magnitude key effect
  coefficient size than the full specification.
\item
  33\% of alternative specifications have a smaller magnitude key effect
  coefficient size than the full specification that is statistically
  significant.
\item
  46\% of alternative specifications have a key effect coefficient size
  between full and bivariate specifications.
\item
  84\% of studies have a minimum and maximum possible key effect of the
  same sign.
\item
  The average standard deviation of estimates around the full
  specification across studies is 0.32.
\end{itemize}

\hypertarget{deflators-and-inflators}{%
\subsection{Deflators and Inflators}\label{deflators-and-inflators}}

\begin{itemize}
\item
  Percent of controls that inflate the estimate of the key variable on
  average; unweighted: 51\%, weighted by study: 49\%.
\item
  Percent of controls that statistically significantly inflate the
  estimate of the key variable on average; unweighted: 39\%.
\end{itemize}

~

\includegraphics{covariates_appendix_files/figure-latex/unnamed-chunk-20-1.pdf}

\hypertarget{articles-text-about-controls}{%
\section{Articles' Text about
Controls}\label{articles-text-about-controls}}

This section presents the text about controls from observational
articles that did not present a bivariate specification.

1

Controls: Economic growth: Increasing income per capita is an important
variable in our theoretical model as it increases support for welfare
state spending, and is correlated with wage inequality. The percentage
elderly can increase the generosity of welfare policy platforms, in
particular on pensions, as the proportion of elderly in the electorate
grows (Persson and Tabellini 2000). Demographic changes can also have an
impact on wage inequality (see e.g.~Hagemann and Nicoletti 1989 for a
discussion of the effects of population aging on the labor market).
Trade openness: implies a higher risk of income loss, and thus higher
support for social insurance. Trade openness might also increase wage
inequality (e.g.~Wood 1994). Union density: There is a large literature
on the impact of the unionization of the working class on the generosity
of the welfare state (e.g.~Korpi 1983). Encompassing unions also tend to
equalize wages (Wallerstein 1999). The definition and sources of the
control variables are described in Table A1. All variables are lagged
one year, i.e.~they refer to the situation the year preceding the
election. Columns 1 and 3 present ``stripped-down'' models including the
country fixed effects, the controls for the time trend, and the source
dummies only. Columns 2 and 4 include con- trol variables. Control
variables, and why we account for them in the regressions, are described
in Appendix B in the supporting information.

2

As controls, I use the same variables included in the matching stage. As
a measure of judicial independence, I adopt the data provided by
theCingranelli-RichardsHumanRights Data Project (CIRI; 2009;
JUDICIALINDEPENDENCE), which are coded 0 for not independent, 1 for
partially independent, and 2 for generally independent. I include
ameasure of regime type using the Polity IV data (Marshall and Jaggers
2002; POLITY) because democracies aremore likely to respect human rights
(Davenport 1995, 1999; Poe and Tate 1994; Poe, Tate, and Keith 1999).
Newer regimes and well-established regimes may have different
preferences, so I control for this factor using the Polity IV data
(Regime Durability). Foreign wars and civil wars may result in periods
of increased repression (Fariss and Schnakenberg 2014; Hill and Jones
2014; Schnakenberg and Fariss 2014). Civil wars, in particular, may
result in periods of lawlessness during which even powerful legislatures
have a diminished capacity to con-strain the other branches of
governments. I use data from the UCDP/PRIO armed conflict database. NGOs
play a key role in political mobilization against oppression and may
succeed in improving government practices. I include the number of
international NGOs (INGOS)in a country using the data provided by
Hafner-Burton and Tsutsui (2005). Economic development is a well-known
predictor of human rights practices (Henderson 1991; Poe and Tate 1994;
Poe, Tate, and Keith 1999), and I control for this using a measure of
per capita gross domestic product (GDP) provided by the World Bank. I
use the natural log of this measure because this effect is likely
nonlinear (Davenport 2007a). To address potential differences among
states of different sizes and potential monitoring biases based on this
factor, I follow much of the literature in including the natural log of
a state's population, using data provided by the World Bank.

3

For our first set of analyses, we control for several party and country
attributes that might be argued to in- fluence the extent to which our
hypothesized relation- ships hold. First, for individual parties, we
control for Party age (measured in years), Party size (measured as a
party's share of the national-level vote), and whether or not a party is
a Niche party (i.e., ecologist, communist, or nationalist party). At the
country level, we control for New democracy, the Effective number of
electoral parties (ENEP), and Economic performance.

4

In addition to the above variables of interest, the models include five
sets of variables that measure respon-dents' socioeconomic status,
perceived commonalities with Latinos, group identification, religious
indicators, political knowledge, basic demographics, and contex-tual
indicators. 13 Extant research commonly notes Latino immigration
attitudes are associated with these factors (Binder, Polinard, and
Wrinkle 1997; Branton 2007; Carey, Branton, and Martinez-Ebers 2013;
Citrin et al.~social and economic integration (as measured by commuting
to work) with the urban core (U.S. Census Bureau 2012, 5). 12 This
measure includes only the respondents who lived in an area where at
least one protest occurred before the respondent was surveyed. 1997;
Hood et al.~1997; Rocha et al.~2011; Sanchez 2006). These variables are
important in controlling the demo-graphic and political differences of
respondents between the treatment and control groups. The assignment of
the respondents to the treatment and control groups was al-most random,
yet the respondents' decision toparticipate in the survey might be
influenced by their personal char-acteristics and exposure to the
rallies. Indeed, we find significant differences in the demographic and
political features of the respondents in the treatment and control
groups. Accordingly, these control variables play an im-portant role in
isolating the effects of the rallies on the outcome variable.
Respondents' socioeconomic status is measured by their educational
background, income level, andpersonal financial situation. Education is
measured by an 8-point scale ranging from 0 to 7. Respondents' income is
mea-sured by four variables that equal 1 if respondents belong to one of
the quartiles on the scale of income and 0 other-wise. The lowest
quartile serves as the baseline category. Additionally, we include an
indicator variable that equals 1 if respondents refused to report their
income status and 0 otherwise. Respondents' assessment of their personal
financial situation is measured by a 3-point scale rang-ing from (1)
gotten worse, (2) stays about the same, and (3) gotten better.
Respondents' perceived commonalities with other Latinos is measured by
an item asking, Thinking about issues like job opportunities, education
or income, how much do {[}ethnic subgroup{]} have in common with other
Latinos or Hispanics? Would you say {[}ethnic subgroup{]} share a lot in
common, some things in common, little in common, or nothing in common
with other Latinos? The measure is coded (1) nothing, (2) little, (3)
some, and (4) lot. Tomeasure group identity, we use a survey item that
asks respondents to choose the category that best de-scribes them:
Latino/Hispanic, American, their national-origin group, andnone of
these.We created twomeasure to capture their primary group identity. For
the firstmea-sure, a value of 1 was assigned to respondents who chose
their national-origin group and 0 otherwise. The second measure assigns
a valueof 1 to respondentswho identified more strongly with their
American identity and 0 oth-erwise. The baseline category reflects a
Latino/Hispanic primary identification. To measure a respondent's
political knowledge, we rely on an additive measure based on responses
to three survey items.Respondentswereaskedto identify theparty that had
a majority in the House of Representatives, the presidential candidate
whowon the 2004 election in their state, and the political party that is
more conservative than the other. The responses were coded 1 if a
response was correct and 0 otherwise. Scores range from 0 to 3, with
higher scores indicating a higher degree of political knowledge. To
account for the impact of religion and civic participation, the models
include three measures: Catholic,Attend Church,andCivic
Participation.Catholic is a dichotomous measure that indicates whether a
respondent is Catholic.Attend Churchrepresents regular church
attendance, which is a dichotomous variable coded 1 if a respondent
indicates she attends religious services once a week or more.Civic
Participationis a di-chotomous measure that indicates whether a
respondent is involved in a civic organization. Given the importance of
the Catholic Church and civic organizations in activist efforts---not
only with regard to the 2006 rallies but also more generally (Schmidt et
al.~2008)---we include this to account for the potential influence of
involvement in the church and civic organizations on immigration
attitudes. We control for differences in national origin using dummy
variables for country or region of origin/descent. Thedummyvariables
equal 1 if respondentsdescend from Mexico, Cuba, Puerto Rico, the
Dominican Republic, Central America, and South America. Mexican
respon-dents serve as the baseline category. Further, we control
fordifferences inagemeasuredbyyears andgender,which equals 1 if the
respondent is female and 0 otherwise. Finally, our model includes
aggregate-level measures of the socioeconomic and ethnic context.
Socioeconomic context ismeasured by tract-level percent high school
ed-ucated. Ethnic context is measured by the percentage of Latinos at
the tract level. These measures are included to account for variance in
immigration attitudes attributable to the context in which one resides.

5

Seven control variables capture country- and election-specific factors
that may influence the probability of a democratic transition. First, it
is likely that democratization occurs primarily in elections that
deter-mine who will hold executive power. Because the stakes are higher
in these contests, they are more likely to elicit higher levels of
domestic mobilization for democracy. I include a dichotomous variable,
main election,coded as a 1 for presidential elections in presidential
(or mixed) systems and legislative elections in parliamentary systems
(Simpser and Donno 2012). Second, I include a variable indicating
whether theincumbent was running in the election, which is expected to
decrease the chances of de-mocratization (Cheeseman 2010; Maltz 2007).15
It is also important to account for the country's previous experi-ence
with elections. Lindberg (2006, 2009) argues that holding elections
fosters institutional change and greater respect for civil liberties and
that these changes cumulate over time. Relatedly, Bratton and Van
deWalle (1997) ar-gue that the more elections held under
authoritarianism, the greater the chances for democratization. I thus
in-clude a variable that sums thenumber of previous elections held under
a continuous authoritarian spell. 16 Two variables control for economic
conditions: first avariableforGDP per capita,lagged one year. 17 High
income is reliably associated with democracy, though its effect on
democratic transitions is less clear. Przeworski et al. (2000, chap. 2),
for example, find that dictatorships at high (but not the highest)
levels of income are more likely to democratize. Second, I include a
variable for GDP growth, measured as the percent change in a coun-try's
GDP from yeart-2tot-1.Ifitistruethatgood eco-nomic performance bolsters
authoritarian regimes, the coefficientonthisvariableshouldbenegative

6

Our analyses included a standard set of demographic and political
control variables, as well as additional theo- retically relevant
contextual and individual-level controls. First, we control at the
county level for the Median House- hold Income, Percent Black, Total
Population, and Bush Vote 2004 to account for the distinct effects of
variance in ab- solute economic conditions, racial composition, county
size, and political culture across respondents' county of residence.7 At
the individual level, we include standard controls for Education, Age,
Gender (1 = male), Employ- ment Status (1 = unemployed), Union
Membership (1 = respondent and/or spouse is union member), Party Iden-
tification (5-point scale; 5 = ``strong Republican''), Ide- ology
(5-point scale; 5 = ``very conservative''), and Re- ligious Attendance.
We control for unemployment and union membership as standard controls in
research an- alyzing attitudes toward the economy and government
redistribution (e.g., Anderson and Pontusson 2007; Fong 2001; Scheve and
Slaughter 2004), and for religious at- tendance because religiosity has
been linked to general ideological orientations, such as humanitarianism
and egalitarianism (Bartels 2008; Feldman and Steenbergen 2001) and
social welfare policies (Sears et al.~1997).

7

Alternatively, it is possible that towns with casualties tended to be
environments of lower socioeconomic sta-tus. Though the differentiation
in induction rates across income classes is disputed (Flynn 1993;
Zeitlin, Lutter-man, and Russell 1973), poor and working-class men are
significantly overrepresented in combat-related deaths (Kriner and Shen
2010; Zeitlin, Lutterman, and Russell 1973). Table 4 reports estimates
from probit analyses that include socioeconomic and political contextual
variables as well as interaction terms between each contextual vari-able
and the binary indicator for a low lottery number. The first two columns
include contextual-level socioeco-nomic indicators. Binary variables
indicate whether the town population is above the median for the sample,
the town's median household income is above the sam-ple median, and the
percent of residents who did not graduate from high school is below the
sample median. Columns 3 and 4 report results of probitmodels inwhich
binary indicators for above-median turnout in 1968 and above-median
shares of Democratic vote share in 1972 are added. Results indicate that
accounting for local po-litical or socioeconomic attributes does not
substantively change the estimates of the interaction effect of a son's
as-signment to high draft priority status within the context of a prior
local war casualty.

8

Throughout the analysis, a number of other variables are used as
controls. All are described further in the sup- porting information
(item 1). Intensity of the trafficking problem in countries of origin,
transit countries, and desti- nation countries was generated based on
the 2006 United Nations Trafficking in Persons report and is a constant
for all years. Missing information measures the availabil- ity of
information on trafficking, which may influence the ability of the
United States to include a country in the report in the first place. It
is a count of how often a country has missing information on 10 types of
data in a given year, including seven unrelated to trafficking. To
reflect the U.S. State Department's access to information about
trafficking in the country, we also created a variable, NGO density,
based on the number of NGO mentions in the U.S. TIP reports, extended
backward to all years, creating a constant measure for almost all
countries included in the analysis. Finally, regional density of crimi-
nalization measures the proportion of countries within a country's
region that had criminalized as of the previous year. Other variables
include civil liberties from Freedom House, an indicator of 2000 TIP
protocol ratification, to- tal population (logged), as well as measures
of a country's bureaucratic quality, rule of law, corruption, or the
share of women holding seats in parliament. All sources and mea-
surement details are listed in the supporting information.
\ldots{}\ldots{}\ldots{}.{[}after analysis{]} Other factors also matter.
Countries are more likely to criminalize the greater the share of women
in parliament, the greater their civilliberties, the greater the
regional density of criminalization, and if they have ratified the 2000
TIP protocol. Interestingly, U.S. aid in itself appears to have little
effect on criminalization.

9

These effects are notably resilient to the effect of issue position
extremity. Issue position extremity does significantly increase
thermometer bias, but farmoreweakly than sorting or partisan identity
strength. Controlsare included for education, sex (dummy), white race
(dummy), age, southern residence (dummy), urban residence (dummy),
frequency of church atten-dance (as ameasure of religious commitment),
and evan-gelicalism (as a measure of religious conservatism, a dummy
variable based on denomination). All continu-ous variables are coded to
range from 0 to 1

10

Allmodels control for basic demographics, including
educa-tion,age,andgender.

11

In addition to these model-specific control variables, we include the
number of standing committees to which a bill is referred, which we
expect to have a positive impact on the level of legislative scrutiny,
and thus the number of changes made to government bills. We also control
for the (logged) size of the bill introduced, since bills with many
articles are naturally going to have more articles changed, on average,
than bills with few articles. We also include a measure of the number of
days a bill spends in the leg- islative process, as well as an indicator
for whether a bill expires before the plenary vote. We also include an
indica- tor for the numerical status of the government to account for
the possibility that minority governments may have to make more policy
changes than majority governments to entice opposition parties to
support legislation. Finally, we include indicators to account for
country-specific and issue-specific fixed effects.

12

In Model 2, we control for evaluator (Ideol- ogy of Eval.) and target
ideology (Ideology of Target; both continuous measures), sex, and
average attraction levels.

13

A vector of variables, includes the mean, maximum, and minimum ability
scores of all representaives, who form the entire pool of potential
candidates, and Xjs is a set of group controls as in Equation (5). To
account for the nested na-ture of the data, we include(2) , a random
intercept for regions. we also include the manager's ability and the
followinggroup-level controls (Xjs): number of group members, age of the
group, and a measure of ethnic fractionaliza-tion.26To account for the
nested nature of the data, we include b2 , a randomintercept for
regions;finally, js isthe residual error term. table note : Controls,
centered on their mean values, include the number of association members
(in units of 50), the DC age, and its ethnic homogeneity (ELF) using a
simple Herfindahl concentration index

14

The most serious potential confound for the analysis be-low is
distinguishing relative political importance and relative population.
Similarly sized ethnic groups are thought to be more likely to clash
(Buhaug, Cederman, and Rød 2008; Horowitz 1985; Montalvo and
Reynal-Querol 2005; Reynal-Querol 2002) and are also likely to be
similar in political importance. I control for: Demographic
polarization=n 2 i nj +nin 2 j , (1) where ni and nj are the population
shares of the plurality group in the enclave and the group opposed to
state-hood, respectively. 27 Esteban, Mayoral, and Ray (2012) argue that
cultural distance exacerbates the effects of po-larization. I calculate
cultural polarization as suggested by Fearon (2003): Cultural
polarization= n 2 i nj +nin 2 j dij, (2) where dij is the linguistic
distance between the enclave plurality and the opposing language group,
normalized to fall between 0 and 1. 28 Additional Confounds Other
confounds are variables that may influence political importance to the
Congress and violence. If grievances caused groups to both vote against
the Congress and use violence, relative INC representation might be
cor-related with violence by virtue of proxying for dissatisfaction.
Therefore, I control for the absolute level of Congress representation
of statehood proponents (Ln enclave plurality group's INC rep.). Other
likely correlates of violence plausibly related to political impor-tance
are population (Ln enclave plurality group's population); economic
development, measured as the share of the workforce in agriculture
(Agricultural labor sharin enclave); and distance to the capital (Ln km
to New Delhi). 29 Regional inequalities were of limited salience dur-ing
reorganization because states levy few taxes. How-ever, states do have
authority to tax and redistribute agri-cultural holdings. Demand for
land reform may be an important control, therefore. Landless rate in
enclave is the share of the agricultural workforce that is landless.
Finally, I also measure enclaves' Hindu population share (Hindu share in
enclave). Wilkinson (2008) and Capoccia, Saez, and de Rooij (2012)
suggest that religious disputes in India have been particularly violent.
Brass (1974) argues that partition made New Delhi wary of territorial
demands construed in religious terms. Since 29 Data on population,
sector of employment, landholdings, and religion are from Census of
India (1951). religion and voting patterns are also correlated, religion
is a potential confound. \ldots{} I control for polarization, the
enclave-plurality group's INC representation, and population, all
variables that might be proxied by relative INC representation. Also,
the literature suggests that New Delhi was least accommodating of
religious minorities and movements posing a separatist threat;
therefore, I control for religious composition and distance to New
Delhi.

15

Of course, controlling for potential confounders is important.
Personalism might, for example, covary with geographic factors in a way
that obscures the relationship between regime type and pursuit of
nuclear weapons. Similarly, personalismcouldbe associatedwith economic
development or military capabilities (in the sense of the Correlates of
War material-resources index), something thatmay be associatedwith a
greater likelihood of pursuit of nuclearweapons. Inviewof these
concerns,we estimate logistic regression models that control for
confounding variables, but keep in mind the importance of limiting the
inclusion of ``posttreatment'' variables that would bias our estimates
of the effect of personalist regime type. Since observations over time
within a particular country are clearlynot independent, failure
toaccount for temporal dependencewithineachcross-sectioncan result in
underestimates of standard errors, leading to unduly optimistic
inferences (Beck, Katz, and Tucker 1998). We thus include three
regressors to model time passed with-out the pursuit of nuclear
weapons:t, t 2 ,andt 3 (Carter and Signorino 2010).

16

Verified voter registration status and a full set of demographic
controls are available for 2,249 of the CCES respondents, but the
results are nearly identical when the analysis is expanded to include
all 4,435 respondents who reside in one of these districts.

17

Demographic control variables. We include age be- cause older
respondents may be more likely to hold more negative views of
government. Education is controlled for because the literature suggests
that as levels of educa- tion increase, respondents may possess a higher
sense of efficacy and trust in government. We also include gender to
examine differences between Latinos and Latinas. The second set of
variables is related to genera- tional status, time in the United
States, and national origin groups. The models include a dummy variable
for whether respondents are first generation, since they are less likely
to be politically acculturated and may be more likely to find politics
confusing. Given differences in Latino political attitudes based on
national origin groups (Abrajano and Alvarez 2010; Alvarez and
Garcia-Bedolla 2003), we also include dichotomous measures for whether
respondents are of Mexican, Puerto Rican, Cuban, Do- minican, or El
Salvadoran decent.12 For example, Cubans are usually the most
conservative among national origin groups, whereas Mexicans and Puerto
Ricans are typically the most liberal. We also include a dummy variable
for respondents who completed the instrument in Spanish, Spanish Pref,
in the event that Spanish-dominant speakers vary in their attitudes.
Similarly, we utilize a variable that measures the percent of a
respondent's life spent in the United States.13 Finally, we also include
three variables related to me- dia consumption because the source of
news on poli- tics could shape how respondents viewed the marches and
their relationship with government. We include a measure for how
frequently respondents watch the news on television. Second, we utilize
a question asking how often respondents read a daily newspaper. We
created a dichotomous variable indicating whether people rely

18

The frequency with which a representative speaks on the floor against
the war in Iraq. Most importantly, Republi- cans are significantly less
likely to criticize the war policies of a Republican president than are
Democrats. Moreover, even after Obama's ascent to the White House, after
five years of public support for the war, Republican mem- bers should be
significantly less likely to criticize it than their counterparts across
the aisle. As a result, the model first includes a dummy variable
identifying Republican members of the House. Moreover, to assess whether
the effects of constituency casualties are different for Demo- cratic
and Republican members, the model also includes the interaction of this
Republican dummy variable with the local casualties measure described
previously. A member's position within the chamber hierarchy might also
affect her or his willingness to use the insti- tutional forum of a
floor speech to criticize the war. To control for the possibility that
members of the leadership might be more or less willing to attack the
White House's policies, we include a dummy variable identifying those
holding leadership positions. Alternatively, members of committees
dealing with foreign affairs and intelligence might be more willing to
defend Congress's institutional prerogatives and confront the
executive's handling of mil- itary affairs. To account for this
possibility, we include a count of each representative's memberships on
the for- eign relations, armed services, intelligence, or homeland
security committees. As a final control for institutional context, we
include a measure of each member's seniority within his or her chamber.
More senior members who are more invested in their institution may be
more willing to confront the executive branch in the military arena;
moreover, particularly in the House where floor time is more tightly
regulated, more senior members may simply have more opportunities to
express their opinions on the war than their junior colleagues (e.g.,
Hall 1996). A significant literature on civil-military relations sug-
gests that veterans of the armed forces may view military matters
differently than civilians (Dempsey 2010; Feaver and Kohn 2001; Gelpi
and Feaver 2002). Accordingly, we include an additional variable,
whether each House mem- ber had served in the armed forces, to the
models. Given the importance of members' personal backgrounds in
influencing their voting behavior (Burden 2007), we also included a
series of demographic variables to identify each member's race and
gender (Dodson 2006; Rocca, Sanchez, and Nikora 2010). Finally, because
this first model pools data from multiple Congresses, it also includes
dummy variables for the 109th, 110th, and 111th Congresses.

19

Political Predispositions. We control for several factors. Two questions
commonly included in the American Na-tional Election Studies (ANES)
survey tapping egalitarianism (i.e., worry less about equality and gone
too far with equal rights) were combined to form a weak scale ( = .48, r
= .31). Two standard ANES items tapping individualism (i.e., blame self
if don't get ahead and poor because they don't work hard) were also
combined to form a scale ( = .49, r = .32). Both individualism and
egalitarianism were rescaled to range from 0 to 1 (individualism: M =
0.41, SD = 0.27; egalitarianism: M = 0.55, SD = 0.29). Political
ideology and party identification were measured using the standard ANES
7-point format, recoded to vary from 0 to 1. We estimated all models
with a variable corresponding to the stage of interview, which does not
change the substantive results. The cooperation rate was calculated as
the ratio of completes to completes, partials, and refusals. The overall
response rate was 31\% calculated as the ratio of completes to
completes, partials, refusals, and no answers for numbers that were
clearly households. A maximum of 15 attempts were made at each number.
Interestingly, the two scales were virtually uncorrelated (r = -.02).
This can be attributed to the fact that the items were worded in
different directions; previous research also finds low zero-order
correlations between positive and negative stereotype items in the
absence of statistical corrections for systematic measurement error
(which cannot be performed here due to the small number of items
available in our survey; see Levine, Carmines, and Sniderman 1999).
However, the correlation is more in line with expectations among low
self-monitors (i.e., among respondents scoring below the 25th percentile
of the self-monitoring scale), at r = .13 (for high self-monitors, i.e.,
those scoring above the 75th percentile of the scale, the correlation is
negative, at r =-.18). 1, with higher scores indicating greater
conservatism and identification with the Republican Party (ideology: M =
0.49, SD = 0.32; party identification: M = 0.48, SD = 0.34). We also
control for age, measured in years,4 education (1 = bachelor's degree or
greater, 0 = otherwise), and gender (1 = Female, 0 = Male).

20

All of the models include controls for the incumbent party's vote share
in 2001 (Incumbent Vote 2001), whether the incumbent candidate was
running (Incumbent Candidate), which party was the local incumbent party
(MMD Incumbent,UPND Incumbent), the number of candidates for the
parliamentary seat (Number of Candidates), and the incumbent candidate's
years of experience (Years Since Incumbent First Elected).

21

Why do some individuals engage in more religious activity than others?
And how does this religious activity influence their economic attitudes?
We present a formal model in which individuals derive utility from both
secular and religious sources. Our model, which incorporates both
demand-side and supply-side explanations of religion, is unusual in that
it endogenizes both an individual's religious participation and her
preferences over economic policy. Using data on over 70 countries from
the pooled World Values Survey, we find that religious participation
declines with societal development, an individual's ability to produce
secular goods, and state regulations on religion, but that it increases
with inequality. We also find that religious participation increases
economic conservatism among the poor but decreases it among the rich.
Our analysis has important insights for the debate about secularization
theory and challenges conventional wisdom regarding the relationship
between religious participation and economic conservatism.

22

The likelihood that Congress accommodates the president's requests
depends upon more than just the presence of peace or war. Most
importantly, perhaps, it depends upon just how much money the president
requests. At the margin, we expect that Congress will look more
favorably upon smaller requests than larger ones.
Wethereforecontrolfortheloggedvalueofthepresident's proposal for each
agency in each year. Congress's response to the president surely also
depends upon the level of political support that he enjoys within its
chambers. Presidents who confront congresses with large numbers of
ideological or partisan supporters are likely to secure appropriations
that more closely approximate their requests than presidents who face
off against congresses dominated by the opposition party. Following
Kiewiet and McCubbins (1985a, 1985b), we therefore control for the
percent of House seats held by the president's party in each year. We
also include three economic indicators: the average unemployment rate
during the year when appropriationsareproposedandset;thenationalgrowth
rate since the previous year; and the total budget deficit
fromthepreviousyear.Onemightexpectthatpresidents receive greater popular
support when the economy is
doingwell,andfurther,thattheeconomymightdobetter in times of war due to
increased government spending. By controlling for these three economic
indicators, we preclude their ability to bias the effect of war on
presidential bargaining success. All of our statistical models include
fixed effects that account for all observable and unobservable
timeinvariantcharacteristicsofindividualagenciesandpresidents.

23

Prior work shows that civilian killing in one period is de-pendent on
the dynamics that allowed for killing in the previous period (Eck
andHultman 2007). We includeAll One-Sided Violence(t-1), Rebel One-Sided
Violence (t-1) ,and GovernmentOne-SidedViolence(t-1)tocorrespond to
their respectivedependent variables.Thesevariables shouldex-hibit a
positive relationship with civilian killing. 3 We also expect that
asbattlefieldviolence increases, hostilities spill over into the
population (Downes 2008; Hultman 2007). We consequently control forAll
BattleDeaths, Rebel Battle Deaths,andGovernment
BattleDeaths,themonthlynum-ber of battlefield deaths incurred by all
factions, rebels, and regime forces, respectively. We control for the
source of conflict between the government and rebels to de-termine if it
is significantly related to the targeting of noncombatants.Government
Conflictis a dichotomous variable that uses the UCDP/PRIO delineation of
civil wars fought over territorial (0) or government (1) con-trol.We
also expect longer wars to offer greater incentives for factions to
victimize civilians. As a war wears on and neither belligerent is able
to subjugate the other, factions may turn to victimization to tip the
balance.Conflict Du-rationis the number of months since the beginning of
a conflict episode. Finally, a larger populationoffers greater
opportunities for civilianmistreatment. We
includePop-ulationtorecordeachwar country's yearlypopulationsize
according to the Composite Index of National Material Capabilities data
(Singer, Bremer, and Stuckey 1972)

24

The main controls in this study are similar to those used in previous
studies of state repression (e.g., Davenport and Armstrong 2004; Poe and
Tate 1994): regime type, economic development, population size, ongoing
armed civil conflict (intrastate), and levels of dissent.We also use
different indicators for some of these dimensions for robustness checks.
Each control variable is discussed briefly below. First, it is a
well-established finding in the literature that democracy is positively
related to respect for personal integrity rights (e.g., Bueno deMesquita
et al.~2005; Davenport 1995, 2007a; Hibbs 1973) and hence lower levels
of state repression. Using the Polity database (Marshall and Jaggers
2005), Davenport and Armstrong (2004) established that there is an
important threshold effect for when democracy matters for curbing
repression: on and below a specific high value of democracy (7 on the
Polity measure), there is no impact; above this value, there is a strong
and negative influence in two distinct phases (one exerted at levels 8
and 9 as well as one exerted at level 10). Based on this, we follow the
convention from Davenport and Armstrong (2004) and include a dummy for
Polity scores 8--9 and a dummy for a Polity score of 10. These dummies
have been found to be significantly different from the Polity scores
below 8 in predicting repression and also statistically significantly
different from each other (Davenport and Armstrong 2004). Other
structural characteristics of states have also been found to predict
repression, such as population size and development (e.g., Mitchell and
McCormick 1998). We therefore control for total population size (natural
log) and development measured by GDP per capita using data from
PennWorld Tables, both measures from Urdal (2006). Previous studies have
found that ongoing conflict activity may increase the risk that leaders
will repress their citizens (Davenport and Armstrong 2004; Landman 2005;
Poe 2004).We therefore include a dummy variable for ongoing intrastate
armed conflict, using the Uppsala and PRIO Armed Conflict dataset
(Gleditsch et al.~2002). A related finding is that dissent could
increase repressive action (Davenport 1995; Hibbs 1973). Hence, we
control for the annual number of antigovernment protests, riots, or
strikes involving more than 100 persons from Banks (2002) followingWood
(2008).Thismeasure varies between 0 and 46 in our sample.28 Ina data
structure like the one used here, there is likely no independence
between all observations, andwe should expect that previous repression
levels within a country matter for currently observed levels. Several
studies find that a lagged dependent variable (LDV) of repression is
highly significant (Davenport 2007a; Poe 2004), and we therefore run
statistical models with LDVs, indicating the level of repression in the
previous year (LDV = 2, LDV = 3, LDV = 4, LDV = 5).

25

Granted, randomassignment ofmunicipalities to different levels of BFP
coverage would be the best way to assess any the program's electoral
effects. However, there were no pilots, and the program was phased in
rather quickly over the whole country. As a next best alternative, I
employ a generalized propensity score (GPS) matching approach to attempt
to hold fixed the connection between development and CCT coverage (Imai
and van Dyk 2004) and focus on the independent contribution thatCCTsmake
to incumbent candidate vote share by contrastingmunicipalities that
differ with respect to CCT coverage but that have similar levels of
development, size, political background, and other observable
covariates. This matching procedure requires stratifying observations
into similar groups. The important assumption here is that within
groups, variations in coverage are as good as random. Following (Imai
2004), this was done by computing propensity scores---the treatment
levels predicted by pretreatment covariates---and then partitioning the
data into strata where all observations have similar propensity scores.
The treatment effect is calculated within each strata by a simple linear
regression of incumbent vote share on the treatment variable (CCT
coverage), controlling for the propensity score itself and growth rates.
GPS matching does not overcome the possibility of omitted variable bias.
Nonetheless, there are significant advantages to this approach over
simply estimating an OLS regression. At a minimum, the matching
procedure ensures common support (i.e., that only similarmunicipalities
are actually compared) and relaxes the assumption that causal effects
are the same across all types of municipalities (i.e., it allows for
heterogeneous treatment effects). These advantages are not trivial, as
they make matching procedures less model dependent than OLS regressions.
Moreover, the combination of both approaches, as employed here, allows
for even greater robustness; regression serves as a correction for
potentially faulty matching (Morgan and Winship 2007, 156), and matching
serves as a protection from model misspecification (Ho et al.~2007).
Matching will be a better statistical fix for nonrandom assignment the
better the treatment can be predicted from observed covariates (Morgan
and Winship 2007, 114). This is particularly relevant in the present
case because there are strong theoretical and empirical reasons to
believe the specific propensity score regression employed can ensure
proper identification ofCCTeffects. Much of the strength of the
empirical strategy relies on the inclusion of the government's official
target of coverage for each municipality in the propensity score
regression. This target was computed once by the government, based on
the 2006 national household survey. Even though the target did not exist
in 2002, it provides a reasonably neutral assessment of ``need'' in
eachmunicipality because it is based on social indicators that move
slowly over time. Besides the target, the most powerful predictor of
treatment is precisely the HDI-M, to which I added squared and cubed
terms to improve the fit of the propensity score regression at extreme
levels of coverage, as well as the same control variables that were
included in the OLS models. State fixed effects were added to account
for variation not captured by the substantive variables and only
marginally improved the fit of the regressions. In the end, the
regressions predicting treatment in each year had R2s of at least 0.8
and exhibited good fit across the full range of the treatment values. As
a result, coverage effects are identified based on small deviations from
the predicted levels of coverage, and it is reasonable to treat
variations in CCT coverage within each strata as random.

26

We control for a range of other factors that might influence renewables
growth. First, we include the previous share of renewables in
electricity generation. Our model highlights the importance of
endogenous growth in the renewable sector after exogenous shocks. We
thus ex-pect that past generation capacity is a positive predictor of
future growth. To smooth the data, we use a three-year average of past
renewable shares (fromt?3tot?1). Per capita income and GDP growth are
included be-cause wealth allows countries to invest in clean energy
(Grossman and Krueger 1995). We also include the share of government
expenditures of GDP. This accounts for the size of the government and
thus proxies for public involvement in the national economy. Similarly,
we control for the share of investment in a country's GDP to proxy for
the country's general tendency to invest in pro-duction capacity. Trade
openness, defined as the ratio of exports and imports to GDP, is also
included because pre-vious research suggests that export industry
creation is a key factor in clean-energy policy (Lewis and Wiser 2007;
Lund 2009). All are measured in constant dollars from the Penn Tables
(Heston, Summers, and Aten 2009). Finally, we also control for past de
facto energy policy decisions by including the share of electricity
produced by nuclear plants and hydro installations, as recorded by the
United States Energy Information Administration. 28 An increase in oil
price (our exogenous shock) may lead to an increase in such sources of
electricity instead of the development of renewable sources. For
instance, France's lack of renewable electricity production could be the
result of reliance on nuclear power. Similarly, we include a measure for
the generation share of conventional thermal electricity oil, natural
gas, and coal used in power generation -- to account for a government's
political cost of going against the interests of the fossil fuel
industry.

27

To deal with temporal dependence of recurrent conflict, the data are
structured as a binary time-series cross-section, and I include a
measure of the number of years since a civil war (onset or incidence)
with cubic splines (Beck, Katz, and Tucker 1998). 26 Thus, the anal-ysis
accounts for the amount of time since the last civil conflict broke out
or was active. In the models presented here, I control for factors
likely to influence both the extent of division in the SD movement and
the onset of civil war. These include previous concessions to the
movement, whether the host state is a democracy, and whether the
movement has ge-ographically close kin. Concessions to SD movements
suggest that the state is actively attempting tomanage the SDmovement's
demands andmay decrease the chance of an armed challenge. In addition,
concessions may satisfy some factions' demands and lead themto exit the
dispute. Open competition in democratic states and the norm of
respecting citizen demands could lead to SD movements havingmore
factions. Additionally, democracies are gen-erally expected to be less
likely to experience civil war. 28 Movements with kin in an adjoining
state may be more likely to form factions linked to these kin who seek
to influence politics in their homeland. The existence of a
neighboringstatewithethnickinmayalsoaffect the state's or movement's
willingness to use force (Jenne 2006). 29 Table 3 reports the results of
my analyses of civil war onset and civil war incidence.

28

Seven control variables capture country- and election-specific factors
that may influence the probability of a democratic transition. First, it
is likely that democratization occurs primarily in elections that
deter-mine who will hold executive power. Because the stakes are higher
in these contests, they are more likely to elicit higher levels of
domestic mobilization for democracy. I include a dichotomous variable,
main election,coded as a 1 for presidential elections in presidential
(or mixed) systems and legislative elections in parliamentary systems
(Simpser and Donno 2012). Second, I include a variable indicating
whether theincumbent was running in the election, which is expected to
decrease the chances of de-mocratization (Cheeseman 2010; Maltz 2007).15
It is also important to account for the country's previous experi-ence
with elections. Lindberg (2006, 2009) argues that holding elections
fosters institutional change and greater respect for civil liberties and
that these changes cumulate over time. Relatedly, Bratton and Van
deWalle (1997) ar-gue that the more elections held under
authoritarianism, the greater the chances for democratization. I thus
in-clude a variable that sums thenumber of previous elections held under
a continuous authoritarian spell. 16 Two variables control for economic
conditions: first avariableforGDP per capita,lagged one year. 17 High
income is reliably associated with democracy, though its effect on
democratic transitions is less clear. Przeworski et al. (2000, chap. 2),
for example, find that dictatorships at high (but not the highest)
levels of income are more likely to democratize. Second, I include a
variable for GDP growth, measured as the percent change in a country's
GDP from yeart-2tot-1.Ifitistruethatgood eco-nomic performance bolsters
authoritarian regimes, the coefficient on this variable should
benegative.

29

All estimations also include a set of controls. The first is Relative
Wealth. Research suggests that poorer citizens are more likely to be
persuadable by private goods offers than other voters (Dixit and
Londregan 1996). For the very poor, immediate improvements of material
conditions (even a small cash handout or a bag of rice) take priority
over collective goods that come with credible commitment problems
(Desposato 2007; Scott 1977). In new democracies, low trust in
politicians has been found to exacerbate risk aversion of poorer voters,
and low-information environments erect additional barriers for poor
voters to enforce collective goods promises (e.g., Stokes 2000). Given
the difficulties of reliably measuring absolute wealth, we adopted a
measure used by the Afrobarometer. This measure asks respondents to
evaluate their personal economic situation relative to other citizens
and provides fairly reliable information about the relative economic
condition of the respondent. The measure has five values and ranges from
much poorer than average to much better off than average. Partisanship:
Partisanship is expected to make voters less likely to swing vote. It is
measured as a dichotomous variable that is coded as 1 if the respondent
indicated being active in a political party. Voted forWinning MP:
Whether or not the respondent voted for the incumbent MP in the last
election is used to condition the hypothesized effects of performance
evaluations as indicated above, but we also expect it to have an
independent effect. In the first instance, supporting a candidate who
ultimately won may bias evaluations positively. In the second instance,
one can expect that bad performance can push voters away from their
candidate who won the previous elections. For challenger voters, poor
performance may make them swing towards another candidate but certainly
not the one in power who did a bad job of delivering goods. This is
particularly important in Ghana, where MPs frequently report that
constituents hold them responsible as agents of development (Lindberg
2010), and because all MPs have equally sized discretionary constituency
development funds. Respondents who supported the candidate who won the
2004 elections are coded as 1 and all other respondents (voters for
losers and nonvoters) as 0. Male: In the control for gender, men are
coded 1 and womenare assigned 0.Research suggests thatwomenhave higher
levels of risk aversion in political and economic activities (Eckel and
Grossman 2008), and we thus expect them to exhibit more stable voting
behavior. Age: The controls for the age of a subject-group's cohorts are
18--22, 23--35, 36--55, and 55 and older.16 We expect older individuals
to have more entrenched voting habits and consequently are less likely
to change them (Franklin 2004). Education: To capture formal education,
the subjects' highest level of schooling is included. This ordinal
variable with five levels ranges from no formal schooling to
post-secondary/university education. Better-educated voters will have
more adept reasoning skills and, all else equal, are more critical and
evaluative. Hence, we expect clientelistic swing voting to be negatively
associated with level of education while the intuition is that its
relationship with policy swing voting should run in the opposite
direction. Information: We expect that more informed voters will have
greater confidence in their own political views and consequently are
less likely to swing vote. An index capturing a subject's exposure to
newsmedia based on the frequency with which she gets information from
radio, newspapers, and television is included. Ethnicity: Two dummy
variables derived from a question that asks respondents to identify
their tribe,with 1 in both cases indicating being Ashanti or Ewe,
respectively, were employed. Safe Havens: In safe havens, a single
voter's ballot impact on electoral outcomes is close to 0, and thus the
appeal of switching parties to obtain some end should be marginal.19
Constituency competitiveness is measured with a dummy variable where
safe havens are coded as 1 when one party has won the last several
elections with a margin of victory exceeding 20\%. Other constituencies
are coded as 0. Other measurement strategies and the robustness of the
findings using other approaches towards this variable are detailed in
online Appendix E.6.

30

As a robustness test, Models 1--6 in Table 1 omit all control variables,
while Models 7--15 present the fully specified models. The results,
though not identical, are quite similar, suggesting that the rela-
tionships reported below are not artifacts of model spec- ification.
Consequently, I proceed more confidently to interpreting the fully
specified model.

31

Several theoretically relevant contextual- and individual-level controls
were included in the analysis. First, prior research suggests that the
economic and po-litical environment surrounding citizens may both exert
distinct influences on their attitudes toward immigration (Campbell,
Wong, and Citrin 2006). Data from the Bu-reau of Labor Statistics were
utilized to create a measure of the unemployment rate in each
respondent's county in 2005. The resulting variable is coded to range
from low to high county unemployment. The second contextcommand ual
variable captures the political climate surrounding respondents by
measuring the percent of registered vot-ers in a respondent's county
voting for Bush in the 2004 presidential election. 8 Turning to
individual-level controls, all analyses included standard measures of
educa-tion, income, gender (1=Male), age, citizenship status
(dichotomous; 1=born in the United States), employ-ment status
(1=Unemployed), pocketbook economic evaluations (1=experiencing
financial distress), party identification (standard 7-point scale;
7=strong Repub-lican), and ideological self-identification (11-point
scale; 11=very Conservative). Beyond these standard controls, several
additional individual-level factors have been identified in the
liter-ature for shaping general attitudes toward immigrants. Of these,
prejudice (Citrin et al.~1997; Huddy and Sears 1995) and the strength of
national identity (Sides and Citrin 2007; Sniderman, Hagendoorn, and
Prior 2004) stand out as likely predictors of both immigration threat
perceptions and policy preferences. All analyses include an 11-category
measure of general negative affect toward Hispanics,Hispanic Affect,
with high values indicating strong dislike for Hispanics. Given the
present critique of cultural threat and the argument that residing in an
acculturating context should serve as a tangible source of cultural
threat that is separate from identity-oriented 8 These data were
retrieved from the CNN 2004 Election Results cite listing vote results
by county and state. For information, see
\url{http://www.cnn.com/ELECTION/2004/}. 382 BENJAMIN J. NEWMAN
concerns, controlling for national identity is essential. A measure of
the strength ofNational Identitywas included in all analyses (1=strong
national identity). In addition to prejudice and national identity,
research has demon-strated that personality traits, such as
authoritarianism, can influence
threatperceptionsandopiniononimmi-gration (Hetherington and Weiler
2009); all models in-clude a control forRight-Wing
Authoritarianism.Last,in-tergroup contact theory suggests that having
friends who are immigrants may reduce threat perceptions and in-crease
support for permissive policy positions. To control for this
possibility, all analyses included a dichotomous measure of whether or
not respondents report having any close friends who are recent
immigrants (1=has immi-grant friends). For ease of interpretation, all
contextual-and individual-level independent variables were recoded to
range from 0 to 1.

32

In order to test our hypotheses, we begin by estimating a series of OLS
regressions where the dependent variable is a member's Legislative
Effectiveness Score. Since our hypotheses concern the difference between
women in the majority and minority parties, we include indicator
variables for whether a legislator is Female, and either a Majority
Party Female or a Minority Party Female. A Lagged Effectiveness Score is
incorporated into the analysis to control for the fact that members are
expected to have consistent interest and innate abilities from one
Congress to the next. Seniority and its squared value measure the number
of terms that the member has served in Congress to capture the
institutional influence that might be acquired by more senior members
(and the squared value allows the seniority effect to taper off). While
seniority is relevant to any investigation of legislative effectiveness,
it is especially important to consider in the context of gender and
politics, as it was not until the 109th Congress that women made up more
than 15\% of the House. Therefore, many female legislators have fewer
years of experience than their male counterparts, which may be related
to their abilities to be effective lawmakers. State Legislative
Experience is a dummy variable that captures whether a member served in
the state legislature prior to entering Congress. As Carroll points out,
``many of the women who run for Congress have gained experience and
visibility in state government before seeking federal office'' (2004,
6). In fact, over 40\% of the female representatives in the 107th
Congress had served in their state's legislative body (Carroll 2004, 6),
which one might expect would translate into increased effectiveness.
Because state legislatures vary significantly in their professionalism,
we also interact State Legislative Experience with an updated version of
Squire's (1992) Legislative Professionalism measure to account for the
possibility that members who served in more professional state
legislatures will be more effective in Congress. Majority Party is a
dummy variable for whether a member is in the majority party, which is
thought to be important for policy advancement generally. Majority Party
Leadership accounts for whether a member is among the leadership
(majority party leader, deputy leader, whip, and deputy whip), with a
similar variable included also for Minority Party Leadership. Speaker is
a dummy variable for the Speaker of the House; Committee Chair captures
whether a member is a chair of a standing committee; and Power Committee
captures whether a member serves on the Rules, Appropriations, or Ways
and Means Committees. All of these variables are particularly relevant
as controls for this analysis, as female legislators have been generally
less likely to attain these positions of influence, and we are
interested in women's effectiveness when accounting for these
institutional differences. Distance from Median captures the absolute
distance between the member and the chamber median on the DW-NOMINATE
ideological scale (Poole and Rosenthal 1997) to control for the
possibility of more centrist members offering proposals that are more
likely to find their way into law. Since previous research has
demonstrated that female lawmakers are more liberal than their male
counterparts, especially (until recently) when in the Republican Party
(e.g., Burrell 1994; Frederick 2010; Swers 2005), this variable is
particularly relevant to our study.8 Members' personal characteristics,
including African American and Latino, are incorporated because they
have been shown to be important in earlier studies of effectiveness
(e.g., Griffin and Keane 2011; Rocca and Sanchez 2008). Size of
Congressional Delegation within the member's state captures the
possibility of natural coalitions among members from the same state.
Vote Share and its square are included to allow for the possibility that
members from safe seats can dedicate greater time and effort to internal
legislative effectiveness rather than external electioneering and to
allow this effect to be nonlinear.

\bibliography{covariates.bib}


\end{document}
